[{"data":1,"prerenderedAt":184},["ShallowReactive",2],{"docs-page-cn-/cn/usecase/blog/persistent_agent_memory_next_task":3,"surround-cn-/cn/usecase/blog/persistent_agent_memory_next_task":177},{"id":4,"title":5,"avatar":6,"banner":6,"body":7,"category":6,"desc":6,"description":16,"extension":171,"lead":6,"links":6,"meta":172,"navigation":6,"path":173,"seo":174,"stem":175,"__hash__":176},"docs/cn/usecase/blog/persistent_agent_memory_next_task.md","持久化 Agent 记忆，应当让下一次任务更轻松",null,{"type":8,"value":9,"toc":164},"minimark",[10,17,20,23,26,30,33,36,47,49,52,70,82,106,118,120,123,126,143,152,161],[11,12,13],"p",{},[14,15,16],"em",{},"记忆张量 MemTensor 如何看待连续性、纠错与边界",[11,18,19],{},"假设团队周二发现，只有先启动数据库，集成测试才能通过。周五，另一个 Agent 接手相关任务，需要找到已经确认的准备步骤，以及它适用的条件。团队调整步骤后，保存的说明也需要同步更新。",[11,21,22],{},"这就是我们希望通过 MemOS 实现的体验：有用的知识能够延续到后续任务，并随时接受纠正。",[24,25],"br",{},[27,28,29],"h2",{"id":29},"大上下文仍然需要记忆策略",[11,31,32],{},"更大的上下文窗口让 Agent 能阅读更多材料。应用仍需决定哪些值得保留、谁可以复用，以及怎样处理更新。",[11,34,35],{},"在测试示例中，失败是一次事件；已经确认的准备顺序则是可复用流程，需要附带适用的环境与条件。",[11,37,38,39,46],{},"MemOS Cloud ",[40,41,45],"a",{"href":42,"rel":43},"https://memos-docs.openmem.net/cn/memos_cloud/introduction/memory_types/",[44],"nofollow","按种类组织记忆","，让这些记录分别发挥作用。事件记忆保留发生时间等细节，技能记录可复用的任务步骤，事实与偏好提供用户背景。应用可以控制生成哪些种类，以及检索时包含哪些种类。",[24,48],{},[27,50,51],{"id":51},"用四步衔接持续工作",[11,53,54,58,59,64,65,69],{},[55,56,57],"strong",{},"记录。"," 通过",[40,60,63],{"href":61,"rel":62},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/add_message/",[44],"添加消息","，MemOS Cloud 从对话中抽取记忆，检查重复与冲突，必要时更新已有记录。使用 ",[66,67,68],"code",{},"info"," 可以附带来源、环境等自定义元数据。对周二的测试，应提交确认准备步骤及适用范围的讨论；元数据也为后续过滤提供字段。",[11,71,72,75,76,81],{},[55,73,74],{},"检索。"," ",[40,77,80],{"href":78,"rel":79},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/search_memory/",[44],"检索记忆","结合关键词与向量检索，按相关性排序并移除重复结果。过滤条件可按标签、时间、Agent 或自定义元数据缩小范围；选择记忆种类可包含技能，返回数量限制则控制结果条数。周五接手任务时，检索相关测试说明，与当前请求一起送入模型。",[11,83,84,87,88,93,94,99,100,105],{},[55,85,86],{},"纠正。"," 假设团队增加了自动启动数据库的脚本。如果知道需要更改的记录，可以通过",[40,89,92],{"href":90,"rel":91},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/update_memory/",[44],"修改记忆","按 ID 编辑；用户也可以用",[40,95,98],{"href":96,"rel":97},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/add_feedback/",[44],"自然语言反馈","描述新流程，由 MemOS 定位并更新相关记忆。对于变化中的事实与偏好，",[40,101,104],{"href":102,"rel":103},"https://memos-docs.openmem.net/cn/memos_cloud/introduction/time_awareness/",[44],"时间感知机制","维护历史版本，并结合查询中的时间线索区分当前信息和早期状态。",[11,107,108,111,112,117],{},[55,109,110],{},"停止使用或删除。"," 过时流程可以通过更新接口归档。",[40,113,116],{"href":114,"rel":115},"https://memos-docs.openmem.net/cn/memos_cloud/mem_operations/delete_memory/",[44],"删除记忆","支持删除指定条目，或清除 API Key 所属项目内某个用户的记忆。删除流程也应覆盖应用缓存或摘要中保存的副本。",[24,119],{},[27,121,122],{"id":122},"共享需要维护者",[11,124,125],{},"把流程共享给多个 Agent 前，先明确谁负责维护，以及哪些用户或项目可以使用。",[11,127,128,129,134,135,138,139,142],{},"MemOS Cloud 的",[40,130,133],{"href":131,"rel":132},"https://memos-docs.openmem.net/cn/memos_cloud/introduction/isolation_filters/",[44],"隔离能力","按项目与用户划分记忆范围。Agent 过滤条件可以进一步筛选某个用户在指定 Agent 下的记录；开启 Agent 独立记忆后，Agent 则拥有跨用户、跨会话的长期记忆。后端应把 ",[66,136,137],{},"user_id"," 映射到已认证用户。",[66,140,141],{},"conversation_id"," 影响检索上下文，不承担访问控制职责。",[11,144,145,146,151],{},"对于共享的环境准备说明，",[40,147,150],{"href":148,"rel":149},"https://memos-docs.openmem.net/cn/memos_cloud/features/knowledge_base/",[44],"知识库","将上传文档转化为可检索记忆，并与用户记忆共同召回。一项任务就能同时参考团队文档和此前交互的背景。",[11,153,154,155,160],{},"在支持的 Agent 宿主中，",[40,156,159],{"href":157,"rel":158},"https://docs.openmem.net/cn/openclaw/local_plugin/",[44],"MemOS Local Plugin","把记忆存入本地 SQLite 数据库，并将记录与检索接入宿主的任务生命周期。任务完成后记录对话、工具调用和反馈，在下一次任务前提供相关上下文。反复出现的有效方法可以沉淀为可复用技能。",[11,162,163],{},"验证时使用三个会话：先确认流程，再由有权限的 Agent 应用，最后修改流程并重新提问。切换到另一个项目，以及删除后，也各重复一次请求。同时检查召回的记忆、实际送入模型的上下文和最终回答，发现过期说明或访问范围错误。",{"title":165,"searchDepth":166,"depth":166,"links":167},"",2,[168,169,170],{"id":29,"depth":166,"text":29},{"id":51,"depth":166,"text":51},{"id":122,"depth":166,"text":122},"md",{},"/cn/usecase/blog/persistent_agent_memory_next_task",{"title":5,"description":16},"cn/usecase/blog/persistent_agent_memory_next_task","KbivSl_eKKxcR3F_LZhCinez75y8LQ6EXm3D0MMqVb8",[178,-1],{"title":179,"path":180,"stem":181,"framework":6,"module":6,"class":182,"target":-1,"active":183,"defaultOpen":183,"children":-1,"description":-1},"如何让错误假设远离 Agent 记忆","/cn/usecase/blog/keep_bad_assumptions_out_of_agent_memory","usecase/blog/keep_bad_assumptions_out_of_agent_memory",[],false,1793163788675]